Interesting and surprising results. You said "the order does switch between runs" but I don't see that in any of the results you posted; there is quite a substantial fluctuation though (11% to 65%).

I tried with something closer to the code in perlperf to which I linked earlier. I have Cygwin on Win10. I also have Perlbrew. I originally ran under 5.32.0; then switched to 5.30.0 to match your version — there were no significant differences in results. As my locale uses '.', not ',', I simply swapped the substitution (y/./,/ instead of y/,/./). Here's the code I used:

#!/usr/bin/env perl

use 5.016;
use warnings;

use Benchmark 'cmpthese';
use POSIX 'locale_h';

say POSIX::setlocale(LC_NUMERIC);
say sprintf('%.2g', 3.14);
say sprintf('%.2g', 3.14) =~ s/\./,/r;
say sprintf('%.2g', 3.14) =~ y/./,/r;

cmpthese 0 => {
    sprintf => sub { my $str = sprintf('%.2g', 3.14); return $str; },
    y_sprintf => sub { my $str = sprintf('%.2g', 3.14) =~ y/./,/r; ret
+urn $str; },
    s_sprintf => sub { my $str = sprintf('%.2g', 3.14) =~ s/\./,/r; re
+turn $str; },
};
[download]

Here's a typically result:

$ ./pm_11119934_bench_locale_s_y.pl
en_AU.UTF-8
3.1
3,1
3,1
                Rate s_sprintf y_sprintf   sprintf
s_sprintf  2940825/s        --      -46%      -85%
y_sprintf  5493603/s       87%        --      -72%
sprintf   19713322/s      570%      259%        --
[download]

I ran it 5 times with 5.32.0, then 5 times with 5.30.0. The 87% in the representative run only ranged between 85% and 89% in all runs. So, I'm getting y/// almost twice as fast as s///; compared with your s/// being variably faster than y///.

I added in the 'use strict; use warnings;' which you had. Now the results varied (85% 88% 21% 71% 75%) but 'y_sprintf' was still always faster. (I suspect the 21% is bogus — perhaps some short-lived background process fired up at that time.)

I then added the 'use locale;' that you had and this substantially slowed down processing: 'y_sprintf' was still faster than 's_sprintf' but I wouldn't say significantly so (5% 1% 7% 12% 8%).

I next replaced my subs with your strings (',' and '.' swapped as before). Again, 'y_sprintf' was faster and this time closer to, albeit more spread out than, my original results (75% 89% 87% 92% 90%).

So, I can't replicate your s/// being faster than y///; all my tests indicated that the opposite was the case. I think the addition of 'use strict; use warnings;' was probably unnecessary for the benchmark; and perhaps a distraction. I don't know enough about locales to confidently comment; however, it would seem that if you can live without 'use locale;' that would be a good thing (its doco, locale, has warnings about its use). Perhaps try with some of my variations to see if you get different results.